ArchSutra AI

ArchSutra AI / Reference

ADR-0001: Deterministic CLI plus IDE-native skills

Version 0.1.0 · Source: docs/decisions/ADR-0001-deterministic-cli-plus-ide-skills.md

  • Status: accepted
  • Date: 2026-10-31
  • Deciders: Vatsal Shah

Context#

Users want AI architecture review without managing another API key. Shell scripts cannot call an IDE's authenticated model, and scraping IDE credentials is unsafe and unsupported.

Options#

  1. Server that calls provider APIs (needs keys, costs money, uploads code).
  2. Skills only (no reproducible evidence, hard to test).
  3. Skills for reasoning plus a deterministic CLI for evidence.

Decision#

Option 3. Skills run on the IDE's model through its documented skill mechanism. The CLI produces reproducible, testable evidence and never calls a model.

Consequences#

  • Positive: no extra key, offline, testable, safe by default.
  • Negative: semantic workflows only work inside a supporting IDE; CLI rules are heuristic.
  • Revisit when a host exposes a supported programmatic inference API, or when opt-in server mode is built.